基于双支网络协作的红外弱小目标检测

An Infrared Small Target Detection Method via Dual Network Collaboration

红外弱小目标检测在预警系统和导弹制导中具有重要的作用,一直是红外图像处理中颇受关注的研究方向。由于红外弱小目标具有信杂比低、尺寸小、形状结构不明显和纹理弱等特点,现有的通用目标检测和语义分割网络直接应用到红外弱小目标检测效果不佳,为此提出一种基于双支网络协作的红外弱小目标检测网络(DualNet)。将检测任务划分成两个子任务,即降低漏检和降低虚警,进而设计两个不同的网络架构分别处理,并利用加权融合损失函数将两支网络信息整合,使得DualNet能够有效地平衡漏检率和虚警率。在自建数据集上的实验结果表明:DualNet相较于通用性能较好的FCN、DeepLabv3、cGAN以及U-net语义分割网络模型具备更高的准确率和鲁棒性,其在F1-measure指标上提高了8%;在SIRST公开数据集上的检测性能也显著超过了基于深度学习的红外目标检测模型ACM和MDvsFA-cGAN,以及多个经典的非深度学习红外弱小目标检测方法。研究结果表明,所提出的方法能够有效提高红外弱小目标的检测精度。

Infrared small target detection(ISTD)is a heated topic in infrared image processing,and it is intensively applied in early warning systems and missile guidance.ISTD faces significant challenges such as low signal-to-noise ratio(SNR),small size,lack of distinct shape or structure,and weak texture,making it a demanding task.The performance of conventional object detection networks and semantic segmentation networks considerably deteriorates when applied directly to ISTD tasks.To address this issue,this paper proposes a new dual network collaboration-based image semantic segmentation network for ISTD,termed as DualNet.DualNet divides the task into two sub-tasks,namely reducing missed detections and reducing false alarms,with two sub-networks focusing on their respective targets(with cost reduced)by employing a weighted loss function to integrate sub-network information.DualNet effectively balances the miss detection rate and false alarm rate.Experimental results show that DualNet outperforms general neural network models(e.g.FCN,DeepLabv3,cGAN and U-net)on the ISTD task,with an improved F1-measure by 0.08.Furthermore,our model outperforms ACM and MDvsFA-cGAN,two most representative ISTD models based on deep learning,and several non-deep-learning-based ISTD methods.